Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

IntRR: A Framework for Integrating SID Redistribution and Length Reduction

Zesheng Wang, Longfei Xu, Weidong Deng, Huimin Yan, Kaikui Liu, Xiangxiang Chu

TL;DR

IntRR is a novel framework that integrates objective-aligned SID Redistribution and structural Length Reduction, and dynamically redistributes semantic weights across hierarchical codebook layers by leveraging item-specific Unique IDs (UIDs) as collaborative anchors.

Abstract

Generative Recommendation (GR) has emerged as a transformative paradigm that reformulates the traditional cascade ranking system into a sequence-to-item generation task, facilitated by the use of discrete Semantic IDs (SIDs). However, current SIDs are suboptimal as the indexing objectives (Stage 1) are misaligned with the actual recommendation goals (Stage 2). Since these identifiers remain static (Stage 2), the backbone model lacks the flexibility to adapt them to the evolving complexities of user interactions. Furthermore, the prevailing strategy of flattening hierarchical SIDs into token sequences leads to sequence length inflation, resulting in prohibitive computational overhead and inference latency. To address these challenges, we propose IntRR, a novel framework that integrates objective-aligned SID Redistribution and structural Length Reduction. By leveraging item-specific Unique IDs (UIDs) as collaborative anchors, this approach dynamically redistributes semantic weights across hierarchical codebook layers. Concurrently, IntRR handles the SID hierarchy recursively, eliminating the need to flatten sequences. This ensures a fixed cost of one token per item. Extensive experiments on benchmark datasets demonstrate that IntRR yields substantial improvements over representative generative baselines, achieving superior performance in both recommendation accuracy and efficiency.

IntRR: A Framework for Integrating SID Redistribution and Length Reduction

TL;DR

IntRR is a novel framework that integrates objective-aligned SID Redistribution and structural Length Reduction, and dynamically redistributes semantic weights across hierarchical codebook layers by leveraging item-specific Unique IDs (UIDs) as collaborative anchors.

Abstract

Generative Recommendation (GR) has emerged as a transformative paradigm that reformulates the traditional cascade ranking system into a sequence-to-item generation task, facilitated by the use of discrete Semantic IDs (SIDs). However, current SIDs are suboptimal as the indexing objectives (Stage 1) are misaligned with the actual recommendation goals (Stage 2). Since these identifiers remain static (Stage 2), the backbone model lacks the flexibility to adapt them to the evolving complexities of user interactions. Furthermore, the prevailing strategy of flattening hierarchical SIDs into token sequences leads to sequence length inflation, resulting in prohibitive computational overhead and inference latency. To address these challenges, we propose IntRR, a novel framework that integrates objective-aligned SID Redistribution and structural Length Reduction. By leveraging item-specific Unique IDs (UIDs) as collaborative anchors, this approach dynamically redistributes semantic weights across hierarchical codebook layers. Concurrently, IntRR handles the SID hierarchy recursively, eliminating the need to flatten sequences. This ensures a fixed cost of one token per item. Extensive experiments on benchmark datasets demonstrate that IntRR yields substantial improvements over representative generative baselines, achieving superior performance in both recommendation accuracy and efficiency.
Paper Structure (23 sections, 11 equations, 4 figures, 4 tables)

This paper contains 23 sections, 11 equations, 4 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Generative Recommendation
  4. Item Representation via SIDs
  5. Quantization and SID Construction
  6. Integrating Collaborative Information
  7. Decoding Acceleration Techniques
  8. Methodology
  9. Problem Formulation and Preliminaries
  10. Recursive-Assignment Network
  11. Adaptive SID Redistribution
  12. Length-Reduced Generative Prediction
  13. Training and Optimization
  14. Experiments
  15. Experimental Setup
  16. ...and 8 more sections

Figures (4)

  • Figure 1: Comparison of existing paradigms (gray) and IntRR (blue). Row 1: Decoupled indexing (gray) yields suboptimal SIDs and representation ceilings, as identifiers remain task-agnostic despite incorporating collaborative regularization or UIDs. Row 2: Contrast between baseline sequence inflation ($L$ tokens per item, gray) and IntRR reduction (1 token per item, blue) Row 3: Our RAN achieves objective-aligned redistribution and deep collaborative-semantic integration at generative learning stage.
  • Figure 2: Overall architecture of IntRR. The top pipeline illustrates the whole flow from semantic indexing to generative learning stage. The bottom-left panel details the RAN structure, featuring a dual-mode switch (Teacher Forcing vs. Redistribution). While the bottom-right panel visualizes the semantic alignment and the recommendation learning process, highlighting the joint optimization of alignment loss ($\mathcal{L}_{aln}$) and recommendation loss ($\mathcal{L}_{rec}$) via the balancing weight $\lambda$.
  • Figure 3: Visualization of weight distribution for two items with the same SID. The RAN effectively achieves the redistribution of weights based on collaborative signals, yielding a more refined and unique representation for Item 23 and triggering a distinct semantic shift at level 2 for Item 29.
  • Figure 4: Sensitivity analysis of the weight $\lambda$ of $\mathcal{L}_{aln}$ on Beauty dataset. The performance peaks at $\lambda=0.1$, illustrating the optimal trade-off between hierarchical structural anchoring and collaborative-driven redistribution.